Meituan, better known globally as a Chinese food delivery giant, has released LongCat 2.0: a 1.6-trillion-parameter Mixture-of-Experts language model with a native one-million-token context window. The release lands at an unusual intersection of frontier AI research and hardware geopolitics. The model matches leading proprietary systems on agentic coding benchmarks, and it was pre-trained end-to-end on a 50,000-card cluster of domestic AI ASIC superpods, without a single Nvidia GPU in the training loop.

That combination matters. Until now, most Chinese frontier models leaned on domestic silicon only for inference, while pre-training still ran on foreign accelerators. LongCat 2.0 flips that pattern and delivers a working artifact rather than a demo.

What LongCat 2.0 actually is

LongCat 2.0 is a sparse MoE model with 1.6 trillion total parameters. Only 33B to 56B parameters activate per token, averaging around 48B. That dynamic activation window is the reason a trillion-scale model stays affordable to serve. Pretraining spanned more than 35 trillion tokens over millions of accelerator-days, with no rollbacks and no irrecoverable loss spikes. For a run of that size on relatively young toolchains, that stability claim is the most interesting operational detail.

The model targets one goal above all others: reliable agentic coding. That means code understanding, repository-wide edits, terminal task execution, and multi-step workflows inside agent harnesses like Claude Code, OpenClaw, and Hermes.

LongCat Sparse Attention explained

A one-million-token context window is easy to advertise and hard to run. Standard attention scales quadratically with sequence length, so at 1M tokens the compute and memory bill collapses under its own weight. Meituan tackles this with LongCat Sparse Attention, or LSA, an evolution of DeepSeek Sparse Attention that drops the scaling closer to linear.

LSA layers three orthogonal indexing tricks:

  • Streaming-aware Indexing reshapes how the indexer picks tokens, converting fragmented memory access into contiguous, hardware-aligned reads. The result is coalesced HBM access and much higher effective bandwidth.
  • Cross-Layer Indexing reuses attention saliency across adjacent layers. Because important tokens stay important from one layer to the next, a single indexing pass can serve several consecutive layers at inference time.
  • Hierarchical Indexing applies a coarse-to-fine two-stage scoring scheme. Block-level approximate scoring first shortlists candidates, then fine-grained token selection runs inside that smaller pool.

Each of those strategies can be toggled independently, and they all extend to the three-step Multi-Token Prediction module used for speculative decoding. In the target model, one indexing pass covers two layers. In the draft steps, all three shares a single indexer output.

N-gram embedding and zero-computation experts

Two more design choices keep the parameter budget efficient. LongCat 2.0 inherits N-gram Embedding from LongCat-Flash-Lite, adding 135B parameters in a sparse dimension orthogonal to the MoE experts. With an n-gram size of 5, the embedding space expands more than a hundredfold, capturing dense local token relationships that a pure MoE structure handles poorly. Meituan constrains this module to under ten percent of total parameters, which keeps its advantage over an equivalent expansion of expert capacity intact.

Zero-computation experts take care of the other end of the spectrum. Punctuation, whitespace, and other trivial tokens do not need heavy processing, so they route to an expert that returns them unchanged. A PID controller keeps the average activation inside the target 33B to 56B range. The result is a shortcut-connected MoE backbone (ScMoE) where dense and expert branches run in parallel rather than sequentially.

Training and serving on domestic silicon

The hardware story is the part that will get studied most closely outside the AI community. LongCat 2.0 was trained from scratch on tens of thousands of Chinese-designed AI ASICs, arranged into physical superpods of up to 48 machines each. Inside a superpod, the interconnect is fully meshed and high-bandwidth. Between superpods, traffic runs over RoCE.

Since single-chip memory is smaller than the 80GB on an H800, Meituan extended standard 5D parallelism into a 6D scheme, adding EMBP to parallelise the N-gram Embedding module. ZeRO-1, selective recomputation, and automatic offloading on out-of-memory events round out the memory strategy. The Muon optimiser was deployed at scale with custom kernels for tensor-parallel state deduplication and symmetric matrix multiplication.

For long-context training specifically, Meituan wrote deterministic attention and KL-loss kernels, extended context parallelism beyond 512-way using all-gather, and overlapped LSA top-k indexing with KV all-gather to hide synchronisation cost. The team reports a throughput gain of over 35% versus a naive implementation.

Benchmark results

On agentic coding, the reported numbers are competitive with the leading frontier models:

  • SWE-bench Pro: 59.5, edging GPT-5.5 at 58.6 and sitting alongside Gemini 3.1 Pro and Claude Opus 4.6
  • Terminal-Bench 2.1: 70.8, reflecting stable shell execution and error recovery
  • SWE-bench Multilingual: 77.3
  • BrowseComp: 79.9, and RW-Search: 78.8 for agentic web search

On broader general-agent tests like FORTE and on foundational benchmarks such as IFEval and GPQA-diamond, LongCat 2.0 trails Gemini 3.1 Pro and the Claude Opus 4.x line. The pattern is clear: this is a specialist for software engineering agents rather than a well-rounded generalist.

Independent leaderboard confirmation is still limited, but there is one useful data point. Before the release, the model ran on OpenRouter under the codename Owl Alpha and reportedly processed around 10.1 trillion tokens per month, roughly 559 billion per day. That places it among the most-used models globally by call volume, before anyone knew where it came from.

Post-training with multi-teacher distillation

Rather than training one monolithic generalist, Meituan built three specialist expert groups and fused them into a single student model through Multi-Teacher On-Policy Distillation, which the team calls MOPD. The three specialists cover distinct capabilities:

  • An Agent expert for tool use, API parsing, multi-turn parameter handling, and self-correction to avoid dead loops or repeated calls.
  • A Reasoning expert for mathematics, STEM problem solving, and multi-hop knowledge chains, with adaptive compute based on problem difficulty.
  • An Interaction expert for instruction following, alignment, and hallucination suppression, with clearly bounded safety behaviour that does not sacrifice usefulness.

All three then transfer their capabilities into one unified model, which is what developers actually call at inference time.

Practical use cases

The design points to a specific set of workflows where LongCat 2.0 fits naturally. Whole-repository reasoning is the obvious one: feeding a mid-sized codebase into the 1M-token window and asking the model to trace a bug across many files at once, without the summarisation hacks that shorter windows force. Multi-step terminal tasks inside an agent loop match the Terminal-Bench focus, where the model executes shell commands, reads errors, and retries until a task passes.

Repository-level refactors that span multiple modules and their tests benefit from the same long-context strength. And for polyglot codebases, the SWE-bench Multilingual score suggests the model can port logic between languages while preserving behaviour. All of these run inside standard agent harnesses, so teams do not need to rebuild their tooling.

The bigger shift

LongCat 2.0 is interesting on architecture alone, and the LSA design will likely influence how future long-context models are built. But the more consequential signal is elsewhere. If the full-process-on-domestic-chips claim holds up under independent scrutiny, then the path around export controls has matured from “adequate for inference” to “adequate to train trillion-parameter frontier models.” Whether that changes procurement conversations inside labs and enterprises is a question for the next twelve months, not this release.